Two scans, six weeks apart, and the arithmetic of believing the difference

The Journal has asked four separate imaging physicists the same question over the past year: given the best clinical DXA in routine use, what is the smallest change in appendicular lean mass you would report to a patient as real? The answers clustered between six hundred grams and one and a half kilograms, depending on the machine, the operator, the positioning protocol and whether the two scans were performed on the same device. Nobody said less than half a kilogram. That figure should be printed at the top of every body-composition report and is printed on none of them.

What a DXA scan resolves

Dual-energy X-ray absorptiometry is the reference method in this field for practical rather than theoretical reasons: it is fast, the radiation dose is trivial, it is widely installed, and it reports regional as well as whole-body values. Its coefficient of variation for whole-body lean mass on a well-maintained clinical scanner with a consistent operator is on the order of one per cent, which sounds excellent until it is converted into kilograms. For a person with fifty-five kilograms of lean tissue, a one per cent coefficient of variation implies a least significant change — the smallest difference between two scans that can be distinguished from measurement noise with reasonable confidence — of roughly one and a half kilograms.

Appendicular lean mass, the arms-and-legs subtotal that is the closest DXA proxy for skeletal muscle, has a smaller absolute magnitude and a somewhat larger relative error, and the two effects roughly cancel. Regional values for a single limb are noisier again. None of this is a criticism of the instrument. It is the reason a body-composition report that changes by half a kilogram between visits has told the person nothing, and the reason the trial substudies report group means rather than individual trajectories.

Bioimpedance measures conductivity and calculates everything else

Bioelectrical impedance analysis passes a small alternating current through the body and measures the opposition to it. Lean tissue, being largely water and electrolyte, conducts; fat does not. From the measured impedance, a height term, a weight term and a set of population-derived regression equations, the device produces a fat mass figure. The impedance is measured. The body composition is computed from an equation fitted to somebody else.

The consequences are well documented. Agreement with DXA at the group level is often reasonable; agreement at the individual level is not, with limits of agreement for fat mass frequently spanning several kilograms in either direction, and the disagreement growing at higher body mass index — precisely the population of interest here.¹ Worse for our purposes, the measurement is sensitive to hydration status, recent exercise, recent meals, ambient temperature, skin moisture and time of day, all of which are changing during incretin treatment. A device that reads fat mass as a function of body water, used in a person whose body water is unstable, will report composition changes that are hydration changes. The Journal does not report BIA-derived composition changes from consumer devices, and would not treat them as evidence of anything.

What the substudies were never powered to detect

An imaging substudy inside a large trial is sized to describe rather than to test. The enrolment is set by how many participating sites have a scanner and by what the sponsor budgeted, not by a power calculation against a composition hypothesis, and the analysis is generally pre-specified as exploratory or descriptive. The consequence is that these substudies can report a mean change with a usable confidence interval and cannot support most of the questions asked of them.

They cannot, for instance, establish whether lean-mass change differs between dose arms, because the per-arm enrolment after splitting is in the low tens. They cannot establish whether it differs by age, sex, baseline adiposity or diabetes status, because those subgroups were not enrolled to be comparable. They cannot describe the distribution of individual responses, because the per-participant least significant change is a substantial fraction of the observed mean effect. And they cannot address function at all, because nobody measured it.

Nor was the imaging repeated when the programmes were extended. The two-year semaglutide extension reported weight, waist circumference and cardiometabolic parameters at week 104 and did not repeat the composition substudy, so there is no imaging at all beyond seventy-two weeks in this class.² Whatever the trajectory of lean mass is in year two of treatment, nobody has measured it.

None of this is a scandal; it is the ordinary economics of trial substudies. It becomes a problem only when a descriptive group mean is quoted as though it characterised what will happen to an individual, which is now the normal register of coverage on this subject.

Proportion of loss against absolute kilograms

There is a rhetorical move available to both sides of this argument and it works by choosing a denominator. Report lean mass as a proportion of total body mass and it rises during successful treatment, because fat is falling faster; the treatment looks composition-improving, which it is. Report lean mass in absolute kilograms and it falls; the treatment looks muscle-costing, which it also is. Both statements can be made from the same scan pair without either being false.

The Journal reports both, in that order, and thinks anybody presenting only one should be asked why. The proportional figure is the right one for questions about metabolic quality: a body with a higher lean fraction handles glucose better and carries less ectopic fat. The absolute figure is the right one for questions about function and reserve, because a hip fracture at seventy-eight is not prevented by a favourable ratio.

The two framings also diverge most sharply exactly where the stakes are highest. A person losing twenty-five per cent of their body weight will show an excellent proportional result and the largest absolute lean-mass reduction in the cohort. Selecting the framing selects the conclusion, which is why the trade has settled on whichever one suits it.

The soft-tissue artefact in bone densitometry

Densitometry infers bone mineral density from the differential attenuation of two X-ray energies, using the surrounding soft tissue as the baseline against which bone is distinguished. The algorithm assumes a soft-tissue composition, and that assumption is embedded in the calibration. When the thickness and fat fraction of the tissue overlying a measurement site change substantially, part of the apparent change in bone density is an artefact of the altered baseline.

The magnitude is contested. Phantom and cadaver work suggests errors of the order of one to three per cent for large changes in overlying fat, which is the same order as the real bone changes being reported over a year of rapid weight loss. In practice this means that a hip bone mineral density reduction of two per cent in a person who has lost a fifth of their body weight cannot be cleanly separated into a bone effect and a measurement effect, and the published analyses do not attempt it.

Quantitative computed tomography and high-resolution peripheral imaging are less vulnerable, measure geometry and microarchitecture rather than areal density, and have not been used in any trial in this class. The Journal regards that as the most easily closed gap in the whole body-composition literature.

Two things follow practically and only two. Eating adequate protein and loading the skeleton during rapid weight loss are supported by general physiology, carry negligible risk, and are worth doing. Expecting either to prevent lean-mass loss outright is not supported by anything, and treating a fall in a DXA number as a failure of adherence is a misreading of what the number can tell you.